AI-based predictive tool for identifying druggable mutations in lung cancer using nationwide comprehensive genomic profiling data.

Abstract

e20642 Background: Comprehensive genomic profiling (CGP) plays a pivotal role in precision medicine. However, the low probability of discovering mutation-based treatments, despite the financial and time burden, may prevent eligible patients from undergoing CGP. To enhance the efficiency and efficacy of cancer precision medicine, it is critical to identify patients who are likely to benefit from CGP. This study aims to identify patient characteristics associated with the discovery of mutation-based treatments through CGP and to develop an intuitive AI tool to predict the probability of identifying druggable mutations. Methods: We retrospectively analyzed data from 3,470 lung cancer patients (Cohort 1) who underwent CGP between Jun 2019 and Nov 2023 and were registered in the Center for Cancer Genomics and Advanced Therapeutics (C-CAT) database, which covers 99.7% of CGP performed in Japan. Using clinical information available prior to CGP, we developed an eXtreme Gradient Boosting (XGBoost) model to predict the detection of druggable mutations. SHapley Additive exPlanations (SHAP) was employed to extract features that contribute to the model prediction. Using the identified clinical factors as input, another AI model was built and deployed as a smartphone application to predict the probability of identifying druggable mutations. The app’s performance was tested on clinical data from 1,307 lung cancer patients (Cohort 2) who underwent CGP between Dec 2023 and Nov 2024. Results: The predictive AI model trained on Cohort 1 achieved an AUROC of 0.851 (sensitivity: 0.825, specificity: 0.733). A separate model, excluding patients with at least one druggable mutation detected by small companion diagnostic tests before CGP, showed an AUROC of 0.791. Positive SHAP values were associated with adenocarcinoma and the number of metastatic sites, while negative SHAP values were with male sex and smoking history. These relationships were consistent across tissue and liquid CGP cases. Notably, among tissue CGP cases, patients with lung or bone metastases exhibited a significantly higher rate of druggable mutation detection compared to those without (lung: p < 0.001, bone: p < 0.05). A streamlined AI model was retrained using the most influential clinical factors and deployed as a smartphone application. When tested on Cohort 2, the app demonstrated predictive accuracy with an AUROC of 0.766 (sensitivity: 0.716, specificity: 0.695) and a Brier score of 0.188. Conclusions: This study identified key clinical factors predictive of druggable mutation detection in lung cancer through explainable AI analysis of nationwide CGP data. Based on these findings, a smartphone application was developed to predict the probability of identifying druggable mutations. This app could expand access to targeted therapies by facilitating broader utilization of CGP tests.